Application of the Nonlinear Wave Metric for Image Segmentation in Neural Networks

Abstract:
The application of neural networks and modern machine learning techniques opened up possible applications for image segmentation, where instead of bounding box detection a pixel level segmentation of the input images can be created. Algorithms designed for image segmentation in applications such as medical imaging, surveillance, gesture control, tracking etc. require the definition of a loss function for the comparison between images. While the brain can compare complex objects with ease, the same is usually a very difficult task for algorithm designers. Comparison between objects requires a properly defined metric that determines the distance, similarity between them. In this paper we will show how the application of a topographic metric can increase the accuracy of traditional segmentation algorithms.